BackgroundNeoadjuvant Chemotherapy (NAC) followed by radical cystectomy is the first-line treatment for patients with muscle-invasive bladder cancer. Histomorphologic features from transurethral resection of bladder tumor (TURBT) are currently underutilized in clinical practice. We hypothesize that accurately predicting response to NAC using specimens from TURBT will provide crucial information for better treatment planning and patient stratification. Multiple instance learning (MIL) is an efficient A.I. imaging approach for patient-level or slide-level Whole Slide Image (WSI). Effective feature extraction is a major challenge for WSI, particularly in smaller cohorts. To this end, we propose a novel framework, NAC-AI, that leverages bladder WSIs from TCGA for patch-level tumor detection, a large-scale digital pathology foundational model for feature extraction; and a transformer-based cross-attention MIL model for NAC response prediction.MethodsNAC-AI was developed with a publicly available TCGA dataset, pre-trained foundational model UNI, and in-house TURBT data from patients who received NAC and RC at the University of Washington and were classified as responders (pT0 or pTis) or non-responders (pT2 and/or pN+). The framework included three components: 1) A patch-level tumor detection model using vision transformer (ViT-L16) architecture that pre-trained on TCGA bladder data and fine-tuned on TURBT slides to learn to discriminate tumors against artifacts that were unique to TURBT slides; 2) Slides were patchified, and patch-level features embedding were extracted using UNI, a ViT based digital pathology foundational model, along with tumor probability information from the tumor detection model; 3) The extracted features forming a bag were fed into a cross-attention MIL model built based on TransMIL for the final patient-level binary classification of NAC response. Patch-level probability information served as an extra hard-attention mechanism to guide the model in focusing on tumor regions. State-of-the-art approaches such as Attention-based MIL (ABMIL) and Clustering-constrained-attention MIL (CLAM) with both ResNet50 features and UNI features were compared.ResultsA total of 309 slides (1.04 million 40x patches) from TCGA were used for the path-level tumor detection model. A model with a test ROC-AUC of 0.920 was used for step 2. A balanced 138-slide (70 responders vs 68 non-responders) dataset from the University of Washington was used for tumor detection fine-tuning and NAC-AI MIL model development. Five-fold cross-validation metrics were reported. Compared to ABMIL-ResNet50, CLAM-ResNet50, and CLAM-UNI, our approach achieved a ROC-AUC of 0.70+/-0.07 vs 0.52+/-0.07 vs 0.55+/-0.10 vs 0.65+/-0.12. The corresponding accuracy, sensitivity, specificity, and precision scores are 0.72+/-0.05, 0.81+/-0.9,0.62+/-0.10, and 0.69+/-0.06.ConclusionsOur study addresses the critical clinical need of predicting NAC responses from TURBT specimens. NAC-AI has shown promising correlations between the two, potentially improving treatment planning. Our findings underscore the value of the TURBT specimen after artifact removal using a fully automated algorithm and call for larger cohort and multi-center studies to enhance prediction accuracy.

Learning Objective

1. Understand the potential connection between the TURBT specimen and NAC responses.